A neural network model for tracking marker objects in a video

Marker objects tracking is a key component required for understanding state and position of objects in a video. At present time the most popular approaches for marker objects tracking are those based on neural networks. Despite their excellent performance, it is quite often that they skip or confuse some of the groups of marker objects. In this work we propose a neural network based model that corrects tracking errors of the main model via an analysis of marker objects’ relative positions. The network is lightweight as it takes only 1.6 ms to run on a consumer level GPU making it well suited for real-time applications. It also reduces marker objects tracking error significantly.